Chapter 22: Running It Well#

The rhythm is the system#

The folders and skills are instruments. The rhythm is playing them in time. None of it is heavy — a few commands a day, maybe fifteen minutes total — but run consistently, it is what makes the system compound instead of rot. Here is the loop in the order of a working day.

Morning: /daily-briefing. Pulls your calendar, yesterday’s unfinished items, and overnight messages from whatever you have connected, then proposes a top-three focus for the day — written to the top of today’s Daily Log. Nothing connected yet? It still runs on yesterday’s carry-forwards; it just gets richer as your connectors do.

All day: /brain-dump. Whenever a thought is worth keeping — an idea, a risk, a promise you just made — /brain-dump it in one line. It lands timestamped in today’s raw tape with zero follow-up questions, so it never breaks your flow. This is the keystone habit of the entire system.

Around meetings: /call-prep and /save-meeting. Before a meeting, /call-prep assembles everything the folder knows about the person and the topic into a one-page brief. After one worth keeping, /save-meeting files a structured record.

End of day: /daily-log. Reads the day’s raw tape, sorts it into decisions, to-dos, ideas, and blockers — and refreshes the Active Workstreams block in your CLAUDE.md (Chapter 20), so tomorrow’s sessions start current. Run it twice by accident and it merges instead of duplicating. It never touches the raw tape.

End of week: /weekly-review. Reads the week and gives you six things: decisions made, what shipped, what did not, recurring themes, open threads, and — the point — an opinionated set of priorities for next week.

When an update is due: /stakeholder-update. Gathers from your logs and meetings, then proposes topics and waits for you to pick before drafting. That pause is deliberate. Choosing what the reader needs to hear is the part only you can do; the skill refuses to steal it.

A real day, for shape: briefing over coffee surfaces three meetings and one overnight escalation. /call-prep before the 10am. Three brain dumps through the morning — a pricing idea, a risk, a promised follow-up. /save-meeting after the 10am, /decision-log after the pricing call. /daily-log at six. Six commands, nothing fell through.

When the rhythm breaks#

You will miss days. The system is forgiving by design: everything is date-stamped, so you pick up where you are, not where you should have been. Missed the end-of-day sort? Run it tomorrow. Chaotic week? /weekly-review reconstructs the thread from whatever you did capture. The only habit that genuinely matters is the cheapest one: keep capturing. Briefings and logs can be rebuilt from the tape; the tape cannot be rebuilt from nothing.

Adding tools without sinking the boat#

Every tool you add — every plugin, connector, and skill — is a small permanent tax on every future session, because Claude has to consider all of them every time. The enthusiastic beginner’s failure mode is thirty installed things they cannot remember installing, all quietly slowing every run.

The rule: subtract by default. Build with what you have; install only what you cannot do without. Practical version:

• Before installing, name what it replaces. “Nothing — this is genuinely new” is a green light. “The built-in way, but slightly nicer” deserves a hard look.
• Give every install a two-week trial. Used it in week one, forgot it by week two? Uninstall.
• When a session feels sluggish and crowded, remove tools first. Subtraction fixes more slow sessions than any addition does.

And when you do add: prefer the plugin version when one exists, use a hosted connector when you need a live login to a service, and remember that command-line tools Claude can run itself are the lightest option of all. You do not need to memorize that ranking — just ask Claude “what’s the lightest way to connect this?” and make it justify the answer.

Where to run longer work#

You now know three ways to run work beyond the current conversation, from three different chapters. Here is the one-paragraph router. Short task, you are present, you want to align before anything changes: plan mode (Chapter 4). Longer job with a finish line you can write in one sentence, and you will leave and come back: a named goal (/goal, Chapter 15). Independent job that needs no steering mid-flight: a background agent (Chapter 15). And the honest tiebreaker: if you cannot write the finish line as one checkable sentence, you are not ready to delegate it — stay in plan mode and keep your hands on it.

Trust goes both directions#

Chapter 14 gave you the permission ladder, from ask-me-everything up to full autonomy. The half nobody practices: coming back down. After a near-miss, a confusing failure, or before a sensitive stretch of work, drop back a rung for a week. Climbing as you build trust is obvious; descending when trust takes a dent is the discipline that makes the climbing safe. Decide before you move up what would make you move down — if you cannot name it, you have not earned the rung.

(There is also a sandbox feature that boxes commands in at the operating-system level — /sandbox. I do not use it day to day; if you are curious, the official page is code.claude.com/docs/en/sandbox-environments.)

Which model to run it on#

Run the Command Centre on Opus 5 — the smartest everyday model and the current default. This whole part is thinking-heavy work: drafting, deciding, analyzing. That is exactly where the strongest model earns its keep, and exactly where a mediocre answer costs the most. Sonnet 5 is the fast lane when you want snappier responses on lighter work, and Haiku is for bulk jobs — classify these two hundred rows, rename these files. If what you want is speed rather than a different brain, Opus’s fast mode (/fast) speeds up output without switching models. Chapter 17 has the full cost-and-routing play, including running cheap models inside subagents.

Seven habits that keep it honest#

The short list of operating habits that separate a system that stays good from one that quietly degrades. Each is one sentence on purpose.

1. Two corrections, then write it down — the rule from Chapter 19, and the closest thing this part has to a law.
2. Three failed fixes means stop. When each fix creates a new problem, the plan is wrong; re-plan instead of trying a fourth.
3. Ask for evidence, not assurance. When Claude says done, have it show the file, the count, the passing check. “Looked done” is how things turn out not-done three days later.
4. Find the cause before accepting the fix. If a fix works and nobody can explain why, the problem has moved, not gone.
5. Make sources visible. In any research, have claims labeled by where they came from — a regulatory filing and a forum post are not the same kind of true.
6. Write workarounds where they will be found. The second time a tool trips the same way, the workaround goes into CLAUDE.md or the skill it belongs to — not into your memory of last Tuesday.
7. Promote work one rung at a time. A task you have done twice the same way can become a skill. A skill you run on a schedule can become a routine. A routine that runs without steering can become a background agent. Never promote two rungs at once — and demote anything that keeps needing your judgment back to a hand-run command.

The part nobody can install for you#

The plugin scaffolds the folders and the skills in five minutes. The habits in this chapter are yours to run. Run them, and the system genuinely compounds — month three is startlingly better than week one. Skip them, and the best scaffold in the world rots on schedule. The Command Centre is not something you install. It is something you operate. The installing just gets you to the starting line.